dicast: a machine learning method for accurate structural variant detection from short-read sequencing data Authors: Nico Alavi, M-Hossein Moeinzadeh, Jakob Hertzberg, Uirá Souto Melo, Lion Ward Al Raei, Paolo Infantino, Maryam Ghareghani, Marco Savarese, Stefan Mundlos, Martin Vingron Published: 2026 ABSTRACT Structural variants are a common cause of human diseases, but their detection from short-read sequencing remains challenging, despite being the technology underlying most clinical workflows. We present dicast, a machine-learning method that scores SV calls from short-read data using alignment and genomic-context features. dicast is trained on a new multi-technology ground truth built from nine samples, with extensive manual curation. It outperforms existing short-read callers and consensus approaches, recovering substantially more true positives at high precision. We also demonstrate dicast’s applicability for diagnostics, identifying all pathogenic variants in multiple disease cohorts, and 20% more candidate pathogenic deletions than consensus approaches.Supplementary InformationThe online version contains supplementary material available at https://doi.org/10.1186/s13059-026-04280-y. Background Despite significant advances in our understanding of the human genome, the diagnostic rate for genetic diseases remains low, leaving many patients without a molecular diagnosis [1–4]. Structural variants (SVs) contribute to a wide range of genetic disorders, yet their detection and interpretation remain challenging, limiting their integration into routine clinical diagnostics [5, 6]. Improving SV detection, along with understanding the strengths and limitations of available sequencing technologies, is crucial for increasing diagnostic yield and identifying the genetic basis of previously unresolved cases. When designing diagnostic pipelines, one of the first considerations is choosing the appropriate sequencing technology. Nowadays, there exists a myriad of technologies to read DNA. Short-read sequencing (SRS) typically generates paired-end reads with a size of around 100–150 bp [7]. During linked-read sequencing (LiRS), a short nucleotide-based barcode is attached to each read, indicating which reads originate from the same DNA molecule, providing long-range information that can be utilized for phasing and the detection of SVs [8, 9]. Long-read sequencing (LRS) generates reads ranging from several kilobases to over 100 kb [10]. Additionally, optical genome mapping (OGM) constitutes an orthogonal technology to DNA sequencing where molecular labels are enzymatically added to specific motifs, and the labeled DNA is imaged to detect large structural variations by comparing label patterns to a reference genome [11]. Alongside the choice of sequencing technology, researchers and clinicians must select from an array of computational SV detection methods. Due to limitations in precision and recall of individual detection methods, it is common practice to apply multiple methods on the same patient data and use subsequent integration strategies such as consensus calling [12]. Consensus calling integrates results from multiple detection methods by filtering for variants detected by at least n out of N callers. A higher n improves precision by reducing false positives, while a lower n enhances recall by capturing more true variants at the cost of additional false positives. Due to this inherent tradeoff, better integration strategies are needed, in particular for short-read-based methods, which exhibit high false-positive rates. Machine-learning-based approaches have been proposed to address this tradeoff. Early works such as forestSV trained a random forest classifier on alignment features, and methods like SV2 apply a support-vector classifier to genotype and filter deletions and duplications [13, 14]. However, such methods remain limited in SV-type scope, detection performance, or computational scalability. Accurately evaluating sequencing technologies, SV callers, and integration strategies is non-trivial. Even though there exist several ground-truth SV datasets that these evaluations are based on, they can contain inaccuracies due to shared biases among technologies and methods. Previous efforts to assess SV detection performance were often confined to comparing individual SV callers without attempting to systematically assess the strengths and weaknesses of the underlying sequencing technologies [15, 16]. Furthermore, Khayat et al. demonstrated the existence of notable biases in structural variation detection, particularly in how mapping methods, sequencing centers, and replicates contribute to analytical variability. Their findings highlight the need for multi-sample evaluations [17]. Additionally, in comparison to single nucleotide polymorphisms (SNPs), there is a lack of a definitive method or gold standard to determine the similarity between SVs. Recently proposed tools such as SVCROWS aim to address this gap, for example by introducing size-weighted reciprocal-overlap frameworks that adapt matching stringency to SV size [18]. Together, these challenges highlight that robust evaluation requires the continuous improvement of existing SV benchmark datasets and the generation of novel ones. Following detection and prioritization of variants, a key bottleneck in the diagnostic pipeline is the manual inspection of candidate SVs. This step ensures that each variant call is supported by sufficient sequencing evidence, reducing the risk of false positives. It usually relies on genome browsers such as IGV for visualizing read alignment at variant breakpoints but quickly becomes time-intensive as the number of potential variants grows [19]. Although tools like SVHawkEye and Samplot have been developed to facilitate this process, they either cater to specific sequencing technologies or do not provide sufficient information for comprehensive curation [20, 21]. In this work, we address these challenges by creating a ground truth SV database based on nine samples, leveraging multiple sequencing technologies to ensure a broad and accurate representation of SVs. To streamline the laborious process of manual curation, we introduce a novel technology-agnostic visualization tool, cuban, that enables scalable inspection of SV calls, ultimately supporting the development of our curated SV benchmark set. Using this benchmark, we systematically evaluated five different sequencing technologies and fifteen SV detection methods, assessing their strengths and weaknesses regarding SV type, size, and genomic context across multiple samples. We also explored the benefits of combining multiple technologies to gain a more comprehensive view of the structural variation present in a patient. Given that short-read sequencing remains the workhorse of routine clinical diagnostics, we then developed a novel machine learning model, dicast, which was trained on our ground truth database to learn sequencing patterns of true-positive SVs. dicast outperformed all other tested short-read-based SV callers as well as consensus calling strategies in detecting insertions, deletions, and duplications. Finally, we assessed dicast’s clinical utility in three rare disease cohorts—congenital limb malformations, atrial fibrillation, and neuromuscular disorders—where dicast detected all pathogenic variants while filtering out a large proportion of false-positive artifacts. Results Establishing a high-quality, manually curated benchmark set for SV detection In total, we sequenced nine patient samples employing a comprehensive set of sequencing technologies, including Illumina SRS, PacBio LRS, two linked-read sequencing technologies by 10x Genomics and Universal Sequencing (TELL-Seq) as well as Bionano OGM (Fig. 1A). After preprocessing, we employed a total of 15 different SV detection methods, partly read-alignment based and partly de novo assembly based, to generate a comprehensive set of SV calls across our cohort (Fig. 1B, Additional file 2: Tables S1-2). Across the nine samples, this pipeline produced approximately 35M raw SV calls, dominated by PacBio continuous long read (CLR; ~ 32 M, ~ 93%), with the remainder from Illumina, 10x, TELL-Seq, and BioNano (Fig. 1D). After per-caller quality filtering, approximately 3.9M calls remained, which served as the input to the ground-truth construction process.Fig. 1Establishing a high-quality, manually curated benchmark set for SV detection. A Overview of samples in the ground truth and technologies they were sequenced with. B Workflow for constructing the ground truth set of structural variant calls. C Ground truth construction was performed using an overlap graph. Different types of graph substructures corresponding to confirmed variants, conflicting variants, and singletons are shown. Public databases used in the confirmation process comprise GIAB SV Tier1 v0.6, gnomAD SV v2.1, HGSVC Freeze 3, and Audano EEE SV-Pop. D Sankey diagram of benchmark construction. Raw SV calls per sequencing technology flow through quality filtering, graph-based deduplication into distinct events (≥ 50 bp), confirmation (multi-technology, multi-caller long-read, or public-database support), and manual curation to yield the final benchmark of which a subset of variants were manually curated. Ribbon widths are proportional to variant counts. E Number of variants in the ground truth dataset per sample, stratified by SV type. F SV size distribution for SVs < 1 kb (left) and between 1 and 30 kb (right). The constructed ground truth set is being compared to the HGSVC Freeze 4.0 HG002 benchmark dataset Establishing a high-quality, manually curated benchmark set for SV detection. A Overview of samples in the ground truth and technologies they were sequenced with. B Workflow for constructing the ground truth set of structural variant calls. C Ground truth construction was performed using an overlap graph. Different types of graph substructures corresponding to confirmed variants, conflicting variants, and singletons are shown. Public databases used in the confirmation process comprise GIAB SV Tier1 v0.6, gnomAD SV v2.1, HGSVC Freeze 3, and Audano EEE SV-Pop. D Sankey diagram of benchmark construction. Raw SV calls per sequencing technology flow through quality filtering, graph-based deduplication into distinct events (≥ 50 bp), confirmation (multi-technology, multi-caller long-read, or public-database support), and manual curation to yield the final benchmark of which a subset of variants were manually curated. Ribbon widths are proportional to variant counts. E Number of variants in the ground truth dataset per sample, stratified by SV type. F SV size distribution for SVs < 1 kb (left) and between 1 and 30 kb (right). The constructed ground truth set is being compared to the HGSVC Freeze 4.0 HG002 benchmark dataset To generate a ground truth dataset, we have developed a novel graph-based framework. In this graph, individual SV calls are nodes that are connected via an edge if the breakpoints of two variants from different methods are at most 500 bp apart and their size similarity is at least 70%. Connected components in this graph correspond to structural variants that are supported by multiple methods. The resulting graphs contained on average approximately 64,000 connected components per sample (median size 2, mean 3.18), with 58.9% consisting of exactly two nodes and 41.1% of three or more (Additional file 2: Table S3). We call a variant confirmed if its respective connected component fulfills one of the following three criteria: 1) The variant is supported by at least two methods based on two different sequencing technologies; 2) the variant is supported by at least two long-read-based methods; 3) the variant is supported by evidence from at least one out of four publicly available databases (see Methods). In contrast, we call variants conflicting if their connected component contains two nodes from the same method. Applying these criteria across the cohort yielded approximately 95k SVs confirmed via multi-technology support, 119k via multiple long-read-based methods, and 23k via public-database evidence (Fig. 1D). After creating this first draft of a ground truth dataset, we systematically identified potential errors in our dataset. Variants that were potentially erroneously included or erroneously excluded were manually inspected by a panel of human curators utilizing a novel SV visualization tool called cuban. Given the breakpoints of a candidate SV, cuban creates an image displaying different aspects of the aligned sequencing data to enable human curators to assess the variants' validity. These aspects include coverage (compared to the chromosomal average), the presence of repetitive elements, distributions of insert size outliers and discordant read pairs, and read alignments around the breakpoints (see Additional file 1: Fig. S14 for an example plot and Methods for an explanation of the visualization tracks). We manually curated a total of 11,511 SVs to mitigate potential biases arising from the automated confirmation process. This encompasses the removal of false-positive (FP) variants from the benchmark and the inclusion of previously excluded SVs. The final ground truth SV callset consists of 122,133 insertions, 104,099 deletions, 9,664 duplications, and 335 inversions across all nine samples. On average, each sample possessed 13,570 insertions, 11,567 deletions, 1,074 duplications, and 37 inversions (Fig. 1E). Of the variants in the ground truth dataset, 38% were on average supported by at least two methods from two different sequencing technologies, while 52% were included because they were detected by at least two long-read-based methods. Furthermore, 9% were supported by public databases and 1% were added by identifying false-negative (FN) calls through manual curation (Additional file 1: Fig. S2A, D). The majority of structural variants in our ground truth were supported by at least one long-read-based method (Additional file 1: Fig. S3A). We conducted a comparison between one sample from our newly constructed benchmark dataset—tGenVar—and the published HGSVC Freeze 4.0 benchmark dataset for HG002 [16], focusing on the length distribution and composition of variants. Compared to our per-sample average of approximately 13.5k insertions, 11.5k deletions, 1k duplications, and 37 inversions, the HGSVC benchmark for HG002 contains 13.9k insertions, 9k deletions, 115 inversions, and no separate set of duplication calls. The lower per-sample inversion count in our dataset can be attributed to the absence of Strand-Seq [22] data, a technology specifically suited for inversion detection that was incorporated into the HGSVC dataset. In Fig. 1F, we compared the size distribution of insertions and deletions, distinguishing between small SVs (< 1 kb) and large SVs. The size distribution of deletions was similar between the two datasets; however, our ground truth database exhibited a lower number of insertions. It is worth noting that both databases showed a peak around 350 bp for insertions and deletions, which can be attributed to the presence of Alu elements, a type of transposable element frequently found in the human genome. Additionally, a second peak observed around 1.5 kb is associated with Short Interspersed Nuclear Elements (SINEs). Comprehensive evaluation of SV detection methods and sequencing technologies Leveraging our newly constructed ground truth dataset, we systematically assessed each technology, method, and integration strategy for identifying structural variants across diverse sizes and genomic contexts, with the goal of providing a concise guide for clinicians and researchers. In total, we evaluated 15 SV detection methods across four sequencing and one optical mapping technology. We employed two different caller integration strategies to systematically evaluate the potential of each technology across varying SV sizes and genomic contexts. To ensure robustness, the benchmarking was performed across nine samples. Among the long-read-based methods, debreak and cuteSV consistently outperformed other tools by identifying the highest number of variants across multiple SV types. Specifically, debreak excelled in detecting insertions, deletions, and duplications, while cuteSV showed particular strength in identifying insertions and deletions (Fig. 2A, B; Additional file 1: Fig. S4A, B). Manta was the most effective standalone short-read-based caller, though, like other SRS-based methods, it struggled particularly with detecting insertions (Fig. 2A, B). The use of barcoded linked-reads improved variant detection compared to conventional short-read methods but at the cost of lower precision, especially for duplications and inversions. Optical genome mapping demonstrated moderate effectiveness overall but notably exceeded short-read-based approaches in insertion detection (Fig. 2A, B).Fig. 2Comparative analysis of methods and sequencing technologies for structural variant detection. A Precision and recall for deletion detection. The blue line represents precision, while the orange line represents recall. The confidence intervals and error bars indicates the standard deviation for precision and recall across n = 9 samples for SRS, LRS, OGM; n = 5 for LiRS I; n = 6 for LiRS II. The horizontal bar plots compare the precision and recall when combining multiple methods within one technology. Striped bars describe a more than 2/n caller strategy, while non-striped bars describe a union strategy (Short-Read: Illumina, Long-Read: PacBio CLR, Linked-Read I: 10x Genomics, Linked-Read II: TELL-Seq, Optical Mapping: Bionano). B Precision and recall for insertion detection. C Evaluation of combining multiple sequencing technologies for SV detection. Multiple methods within one technology were combined using the union strategy. Error bars indicate standard deviation across all respective samples Comparative analysis of methods and sequencing technologies for structural variant detection. A Precision and recall for deletion detection. The blue line represents precision, while the orange line represents recall. The confidence intervals and error bars indicates the standard deviation for precision and recall across n = 9 samples for SRS, LRS, OGM; n = 5 for LiRS I; n = 6 for LiRS II. The horizontal bar plots compare the precision and recall when combining multiple methods within one technology. Striped bars describe a more than 2/n caller strategy, while non-striped bars describe a union strategy (Short-Read: Illumina, Long-Read: PacBio CLR, Linked-Read I: 10x Genomics, Linked-Read II: TELL-Seq, Optical Mapping: Bionano). B Precision and recall for insertion detection. C Evaluation of combining multiple sequencing technologies for SV detection. Multiple methods within one technology were combined using the union strategy. Error bars indicate standard deviation across all respective samples To facilitate the comparison of sequencing technologies rather than just individual methods, we employed two caller integration strategies. Specifically, we used a union-based approach (combining variants from all methods based on a particular technology) as well as a more conservative consensus strategy (requiring agreement from multiple callers). The union strategy maximized recall but reduced precision, whereas the consensus approach increased precision significantly but at the expense of recall. Long-read sequencing demonstrated the highest true-positive (TP) detection rates for structural variants, consistently identifying a large proportion of variants, as shown in Fig. 2A, B and Additional file 1: Fig. S4A, B. Precision remained high when combining results from multiple LRS-based callers. In comparison, SRS-based methods were notably less effective, detecting fewer than half of the benchmark deletions and fewer than 10% of insertions. While SRS precision could be increased substantially through consensus strategies, the overall recall was notably lower compared to LRS. The use of barcoded linked-reads, which allow for an additional de novo assembly step before SV detection, led to an increase in detected DELs and insertions but resulted in a decrease in precision compared to conventional SRS methods. Except for a negligible number of variants, linked-read-based methods failed to detect DUPs and INVs. OGM showed modest performance overall but demonstrated superior effectiveness to standard SRS methods in identifying insertions, with consistently high precision. To identify the specific strengths and weaknesses of each sequencing technology in regard to SV detection, we assessed their performance in detecting variants across different size classes and genomic contexts. First, we divided all callsets as well as the benchmark sets into variants of sizes 50 bp—100 bp (small), 100 bp—1 kb (medium), 1 kb—10 kb (large), > 10 kb (very large). SRS-based methods excelled at detecting very large deletions and small duplications but faced limitations detecting insertions larger than 1 kb due to read-length constraints. LRS-based methods demonstrated robust detection for variants up to 10 kb though precision declined for very large variants. Optical genome mapping was especially powerful for detecting very large variants, in particular, surpassing other technologies in detecting very large insertions (Additional file 1: Fig. S4D). In most cases, SV detection relies on the alignment of reads to the reference genome, which can vary in difficulty across different genomic regions. For example, the presence of repetitive sequences complicates the alignment process. Consequently, the ease or difficulty of SV detection varies depending on the specific genomic region, its repeat content, and its structure. Therefore, we investigated context-dependent SV detection performance for each evaluated technology. This was done by comparing the overall performance of a technology with its performance when constraining the evaluation to variants overlapping with certain genomic annotations (see Methods). LRS-based methods are the least impacted by repetitive genomic sequences, reflected by their stable recall across contexts (Additional file 1: Fig. S5A). SRS- and OGM-based methods display a drop of approximately 50% in recall in regions containing variable number tandem repeats (VNTRs), CpG islands, and satellite DNA. In contrast, short and long interspersed nuclear elements, long terminal repeats (LTRs) and short tandem repeats (STRs) have a positive impact on their recall. Precision was most negatively impacted in short-read-based methods, with drops observed in satellite DNA, CpG islands, low complexity regions, and retroposons. To resolve performance within tandem-repeat regions, we further stratified STRs by motif unit length (2 to ≥ 6 bp) and VNTRs by reference-locus length (< 200 bp to > 5 kb, Additional file 1: Fig. S5B). LRS recall remained stable across every class. In contrast, SRS recall collapsed with increasing VNTR length while OGM recall was reduced across all VNTR bins. For STRs, SRS precision declined progressively with motif length, reaching a 39% drop at tetranucleotide and a 33% drop at pentanucleotide motifs. The aggregate VNTR-associated recall drop is therefore driven by long VNTRs and the STR-associated precision loss in SRS is concentrated at mid-length motifs. Lastly, we evaluated potential synergistic effects on SV detection performance when combining methods from two different sequencing technologies. Overall, we did not observe major improvements in recall across all technology combinations and SV types (Fig. 2C and Additional file 1: Fig. S4C). The only modest improvement observed was when combining SRS- and OGM-based methods for detecting duplications. Machine learning reliably distinguishes true variants from artifacts Despite the superior performance of long-read-based SV calling methods, short-read sequencing remains the most commonly used sequencing method in clinical diagnostics, owing to its lower cost, existing infrastructure, and scalability. Yet no single SRS-based caller reaches both high precision and high recall across the full range of SV sizes and types (Fig. 2A, B, Additional file 1: Fig. S4A, B). Conventional consensus strategies reconcile multiple callers by retaining variants supported by at least n out of N methods, with the union (n = 1) and intersection (n = N) as extremes; however, varying n only trades recall for precision in fixed steps and treats every supporting call as equivalent, regardless of the underlying evidence. To address the previously described shortcomings of SRS-based SV detection, we have developed a novel machine-learning-based SV detection method for deletions, insertions, and duplications termed dicast. It takes the same starting pool of candidates as consensus methods, drawn either from unfiltered callsets produced by existing methods or from population catalogs, but replaces the caller-count heuristic with a learned confidence score derived from the alignment and genomic-context features. Concretely, each candidate SV is annotated with more than 100 features computed over multiple bins spanning the variant and its flanking regions. Features include the multi-dimensional alignment patterns of surrounding reads, and the genomic context the variant is located in. These annotations are then fed into an XGBoost model, trained on eight of the nine samples from our newly constructed, manually curated dataset, to distinguish true-positive SVs from false-positive artifacts. To ensure comparability, we evaluated dicast, on an independent public benchmark dataset published by the Genome in a Bottle (GIAB, HG002 Tier 1 v5.0q) consortium [23] as well as on the one held-out sample of our newly developed benchmark. We called SVs using the short-read-based methods delly, manta, lumpy, gridss, and cnvnator. The resulting unfiltered callsets as well as a population catalog of variants, served as input for dicast which subsequently predicted the probability of each variant being present in the sample. We additionally applied SV2, an existing machine-learning-based ensemble caller, which received the unfiltered callsets of the five short-read-based methods as input. For deletions and insertions, each callset was compared to the GIAB benchmark. Since at the time of writing, there does not exist a public benchmark dataset for duplications on GRChg38, we performed the evaluation based on one held-out sample from our newly constructed benchmark dataset. dicast outperformed all six tested SRS-based SV detection methods (Fig. 3A). For deletions, dicast achieves a best-in-class recall of 0.45 while maintaining a precision of 0.85. In contrast, the second-best short- read-based caller, manta, achieves a recall of 0.37 and a comparable precision of 0.92. Furthermore, we detect 45% of TP insertions compared to 10% detected by the second-best method, manta. For duplications, we achieve a recall of 0.27 and a precision of 0.88 using our current model. The second best evaluated short-read-based method, gridss, achieved a recall of 0.23 with a precision of 0.83. When population catalog variants were excluded as input to dicast, overall performance remained largely unchanged. However, recall for insertions declined noticeably, as most candidate variants in this category originated from population catalogs due to the limited capability of the input SV callers to provide candidate SVs (Additional file 1: Fig. S6A).Fig. 3Novel machine learning tool reliably distinguishes true variants from artifacts. A Precision-recall curves for detecting deletions, insertions, and duplications based on the GIAB HG002 Tier 1 v5.0q benchmark (DEL, INS), restricted to its benchmark region BED file and one held-out sample from our newly constructed benchmark dataset (DUP). B Number of TP and FP deletion and duplication calls when using dicast compared to various consensus calling strategies. C Distribution of dicast scores depending on the number of input SV callers, a variant was detected by. D Number of rescued variants, which are usually discarded in commonly employed SV detection pipelines. The first row describes the number of true variants which received a low-quality tag by the respective caller. The second row describes the number of true singletons, variants only detected by one SV caller, which would be filtered out using consensus approaches Novel machine learning tool reliably distinguishes true variants from artifacts. A Precision-recall curves for detecting deletions, insertions, and duplications based on the GIAB HG002 Tier 1 v5.0q benchmark (DEL, INS), restricted to its benchmark region BED file and one held-out sample from our newly constructed benchmark dataset (DUP). B Number of TP and FP deletion and duplication calls when using dicast compared to various consensus calling strategies. C Distribution of dicast scores depending on the number of input SV callers, a variant was detected by. D Number of rescued variants, which are usually discarded in commonly employed SV detection pipelines. The first row describes the number of true variants which received a low-quality tag by the respective caller. The second row describes the number of true singletons, variants only detected by one SV caller, which would be filtered out using consensus approaches Next, we investigated the impact of a variant's size on the detection performance for deletions and duplications (Additional file 1: Fig. S6B). Across all size categories, dicast identified more or a comparable number of TP variants while maintaining an overall lower number of FPs. The largest improvement in terms of correctly identified variants can be observed for SVs between 50–500 bp in size. For larger variants, the primary strength of dicast lies in filtering out FP artifacts while detecting a comparable number of TP SVs. A commonly employed filtering strategy for structural variant calling workflows is consensus calling. We compared various consensus calling strategies against using dicast on the unfiltered callsets of five SRS-based SV detection methods. dicast identifies a comparable number of TP variants as when using the union of 5 short-read-based SV callers with 77% fewer FPs (Fig. 3B). When requiring support from 2/5 callers, the consensus approach identifies fewer FPs SVs. However, this tradeoff comes with a cost of only detecting 46% of the TP variants identified by dicast. The ability of dicast to detect TP variants only called by one or a few methods can be observed in Fig. 3C. While a higher number supporting SV callers corresponds to a higher median dicast score for TP SVs, the median score for SVs only detected by a single method is above 0.7 and above 0.8 for variants detected by two methods. Our methods' ability to rescue commonly filtered out variants is further demonstrated in Fig. 3D. First, for each method, we calculated the number of TP variants that received a low-quality tag by their respective caller and would therefore be filtered out by commonly employed pipelines. dicast assigns a majority of those variants (73%) a high quality score, thus rescuing them. Next, we calculated the number of TP variants that are only detected by an individual method, called singletons, and would therefore be filtered out in consensus calling workflows. dicast rescued 81% of singleton deletions, insertions, and duplications. In the medical field, only a select portion of a patient’s genomic variations have clinical importance. These important variations are typically those located near genes linked to the patient’s specific symptoms or condition. Wagner et al. have compiled a list of challenging medically relevant genes (CMRG) that are difficult to analyze due to their complex or repetitive nature [24]. Alongside, they have published a benchmark set of variants within these genes using sample HG002. We utilized this benchmark to evaluate the ability of commonly used SRS-based SV callers and dicast to detect deletions in CMRG. After filtering for deletions and simple insertions larger than 50 bp the benchmark contained 101 deletions and 57 insertions. Out of those, dicast correctly identified 44 deletions and 10 insertions, whereas the next best short-read-based method, manta, identified 35 deletions and 6 insertions (Additional file 1: Fig. S6C). Currently, dicast is dependent on the candidate variants other methods provide. When providing dicast with a set of predefined regions to check, it correctly identified 91 out of 101 deletions and 30 out of 57 insertions. Beyond predictive performance, we compared dicast's computational footprint to SV2 on the HG002 30x Illumina whole-genome sequencing (WGS) sample. dicast completed feature collection and scoring in 48 min wall-clock time (5.0 CPU-hours across 70 threads) with a peak memory of 2.8 GB, whereas SV2, which can only be run single-threaded, required 21 h wall-clock (21.0 CPU-hours) and 7.5 GB on the same input. Evaluating dicast’s potential to assist in clinical diagnostic workflows Manual inspection of SVs is a critical step in clinical diagnostic workflows. It serves as the last in silico verification of potentially disease-causing genomic variants. Current standard workflows employ IGV or comparable applications visualizing aligned sequencing reads at breakpoint loci, which are then inspected by a qualified clinician or bioinformatician. This practice becomes increasingly challenging with the high throughput from current sequencing technologies—especially when shifting from whole-exome sequencing to WGS with only 1.8% of SVs and InDels occurring in Refseq coding exons [25]. dicast has the potential to automate this process or at least significantly reduce the number of candidate SVs that are to be inspected. Current diagnostic workflows already employ comparable strategies to minimize the number of detected candidate variants before manual curation, such as caller-specific filters and consensus calling [26]. While these strategies can help to reduce the number of false-positives, they are often too rigorous and potentially exclude variants that prove to be clinically relevant. With dicast we aim to optimize this precision-recall trade-off and retain candidate SVs that would be discarded in current diagnostic workflows. To evaluate dicast in the setting of clinical diagnostics, we employed three modified versions of a previously described [27] pipeline for the detection and prioritization of putative pathogenic deletions and duplications centered around a cohort of 18 patients with limb malformations: a short-read pipeline representing the current standard for SV detection that relies on caller-defined quality filters, a second short-read pipeline that includes all (incl. low-quality) calls which were subsequently filtered using dicast, and a pipeline for short- and long-reads with caller-defined quality filters. In each pipeline, we first reduced the initial call-set to rare variants. We then assessed their potential regulatory impact using a comprehensive set of disease-specific annotations, resulting in a set of functionally relevant SVs. A detailed description of this process is provided in the Methods. Briefly, we processed complementary transcriptomic (RNA-seq) and chromatin conformation (Hi-C) data sets for each patient. We then incorporated the processed data in a previously published annotation method and derived overlap-based features for rare SVs to guide the identification of functionally relevant calls [28]. Finally, we identified TP candidates through extensive manual curation, performed on all functionally relevant SVs from each pipeline without any prior filtering. These candidates serve as an unbiased ground truth for our analysis of dicast's performance in clinical diagnostic workflows. We set out to evaluate dicast for three potential applications: 1) Reducing the number of regulatory variants that need to be manually inspected (“callset reduction”), 2) identifying (“rescuing”) low-quality candidates and 3) highlighting novel candidates otherwise only confirmed through additional long-read evidence. For each evaluation we used subsets of the manually curated calls generated by the three pipelines (Fig. 4A). To evaluate dicast with regard to the first application, we applied the method on the variant calls generated by the second (unfiltered Illumina) pipeline and performed the evaluation using all corresponding 230 manually confirmed candidates (Fig. 4B). The sequence context and functional annotation overlap of these candidates are characterized in Additional file 1: Fig. S15. dicast is able to identify the vast majority (96.21%) of manually confirmed deletions as true-positives. We then compared dicast’s performance to a consensus calling approach that considers variants supported by more than one caller as true-positives (Fig. 4C). For deletions dicast clearly outperforms the consensus approach with the average recall of 0.96 compared to 0.74 and only a minor decrease in call-set reduction (82.23% vs. 89.65%) resulting in a median of 29 variants per sample that need to be manually curated. For duplications, both dicast’s and the consensus approaches recall are notably lower (0.45 and 0.48, respectively). However, in this case, dicast would achieve this recall with a considerably higher call-set reduction (97.51% vs. 89.85%). To evaluate dicast regarding the second potential application, i.e., rescuing low-quality calls, we applied the model to the subset of low-quality yet manually confirmed candidates from our second pipeline (unfiltered Illumina) that would be discarded by caller-specific filters (31 out of 230). dicast was able to identify 12 out of 31 while only a single low-quality was supported by the consensus approach. Finally, for the third application, we investigated the proportion of confirmed candidates that would otherwise require the support of a second sequencing technology, i.e., those uniquely shared between the second (unfiltered Illumina) and third pipeline (filtered PacBio and Illumina). In this subset, dicast can identify all manually confirmed variants as true-positives.Fig. 4Evaluating the potential of dicast in the content of clinical diagnostic workflows. A Overlap of manually curated candidates. The figure shows the overlap between the final manually curated sets (“Candidates”) of three pipelines for clinical diagnostics as an unweighted Venn diagram. B dicast predictions for candidate variants. Here we show the frequency of dicast scores for functionally relevant calls grouped as manually confirmed candidates and variants discarded during manual curation. C dicast and consensus performance. This figure compares the reduction in call-set size and recall based on the manually curated candidates when using dicast or a consensus approach. dicast’s recall and call-set reduction for each sample is based on the recommended score threshold of 0.4. D dicast predictions for disease cohorts. In this figure, we display the frequency of dicast scores for functionally relevant deletion and duplications in two rare disease cohorts: Patients with atrial fibrillation (left subfigure) and neuromuscular diseases (NMD) (right subfigure). The absolute number of known pathogenic variants and candidates per dicast score bin are shown as red and orange markers, respectively Evaluating the potential of dicast in the content of clinical diagnostic workflows. A Overlap of manually curated candidates. The figure shows the overlap between the final manually curated sets (“Candidates”) of three pipelines for clinical diagnostics as an unweighted Venn diagram. B dicast predictions for candidate variants. Here we show the frequency of dicast scores for functionally relevant calls grouped as manually confirmed candidates and variants discarded during manual curation. C dicast and consensus performance. This figure compares the reduction in call-set size and recall based on the manually curated candidates when using dicast or a consensus approach. dicast’s recall and call-set reduction for each sample is based on the recommended score threshold of 0.4. D dicast predictions for disease cohorts. In this figure, we display the frequency of dicast scores for functionally relevant deletion and duplications in two rare disease cohorts: Patients with atrial fibrillation (left subfigure) and neuromuscular diseases (NMD) (right subfigure). The absolute number of known pathogenic variants and candidates per dicast score bin are shown as red and orange markers, respectively The candidate call-sets generated using our three pipelines suggest that there is potential in considering low-quality calls in diagnostic workflows. However, harnessing that potential requires manual curation of approximately 5-times as many regulatory variants. This clearly highlights the need for methods like dicast to automate this process and maximize the recall of true candidate variants while minimizing the number of variants to be inspected. Even though the ground truth set of manually curated variants are certain to be present in the affected individuals, their functional impact is yet to be confirmed. To evaluate dicast’s potential on known pathogenic variants (Additional file 2: Table S4) in rare disease cohorts, we applied our short-read pipeline and dicast to an atrial fibrillation (AF) and a neuromuscular disease (NMD) cohort (Fig. 4D). We also employed the Lucid Genome Suite (LGS), a platform for collaborative analysis, for 20 previously unsolved NMD cases [29]. Via this visualization and interpretation framework, we collected experts' opinions on putative pathogenic candidate variants, which were then subjected to downstream validation processes and included in this evaluation of dicast. In both cohorts, dicast can drastically reduce the number of functionally relevant calls with a call-set reduction of 96.70% (AF) and 96.01% (NMD) and correctly classifies all known pathogenic variants as true-positives. Discussion The diagnostic journey for rare disease patients often extends over several years and requires substantial resources before achieving a definitive molecular diagnosis. Even with the dramatic advancements in genetic testing technologies, the current success rates are between 30–40% depending on the disease. This success rate is influenced by several critical factors, including the selection of sequencing technology and the downstream analysis of sequencing data encompassing variant detection, prioritization, and interpretation. The data analysis can further be split based on the investigated variant type. Single nucleotide variants have been studied extensively and are routinely assessed in clinical diagnostic workflows. The role of larger genomic alterations, however, is likely underestimated due to the challenges associated with their detection and interpretation. In our analysis, we therefore focused on SVs in the context of clinical diagnostics, with the aim to overcome current bottlenecks during variant detection and consequently increase the potential for a successful diagnosis. We first created a ground truth database of structural variants consisting of nine samples and five different sequencing technologies. We utilized our newly constructed ground truth to conduct a comprehensive comparison of sequencing technologies, gaining insights into which technologies can be used to maximize SV detection sensitivity. Given the continued use of short-read sequencing due to its scalability and affordability, we developed dicast for short-read-based SV detection, outperforming existing SV callers and consensus calling approaches. Finally, we successfully evaluated our model in several clinical scenarios. The present work is subject to several limitations. The two-technology criterion used to include variants in our ground truth dataset does not account for the individual strengths of specific sequencing technologies. To address this, we also incorporated variants supported by at least two long-read-based methods or those common in the general population. However, this approach may have introduced false positive SVs into the dataset due to shared biases among different detection methods. While we made efforts to mitigate these biases by systematically manually inspecting potential FP and FN variants, this was only feasible for a subset of the final ground truth dataset. More extensive collaborative and distributed efforts in manual curation, e.g., through a public web application, will be crucial to improving the quality of SV truth sets in the future. Furthermore, when evaluating different SV detection methods, we focused exclusively on the presence or absence of variants, without considering their predicted genotypes. However, in clinical settings, the zygosity of a variant is a critical factor. Currently, dicast relies on candidate variants detected by other SV detection methods, limiting its maximum recall to the union of all input callers. To address this, we plan to develop new strategies to expand the set of candidate variants that dicast can evaluate. For instance, candidate loci can be drawn from structural variants discovered in other sequenced samples within the same cohort as implemented by long-read force-callers such as cuteFC [30]. In addition, alignment-free detection could surface candidate loci that alignment-based callers miss, as implemented for long-read data by the recently published method SVDSS [31]. A related limitation concerns the scope of supported SV types. Currently, dicast is restricted to deletions, insertions, and duplications, which constitute the largest fraction of SVs per genome and are the SV classes most frequently assessed in clinical workflows. Inversions and translocations are not yet supported, as their low per-genome frequency yields too few high-confidence positive examples to train a supervised classifier reliably. To address this limitation, substantially larger and collaboratively curated datasets spanning more samples will be required to extend dicast to these additional SV types. Lastly, while dicast currently operates only on SRS data, its framework is not fundamentally SRS-dependent and could in principle be extended to LRS call sets. We deliberately prioritized SRS because this is where SV refinement is most needed since existing LRS callers already achieve high precision and recall across most SV types and contexts (Fig. 2A, B; Additional file 1: Fig. S3A, B). Conclusions With dicast we take an important step towards a more accurate and comprehensive SV detection for short-read data. This has wide-ranging implications for multiple downstream applications. Here, we focused primarily on dicast benefits for clinical diagnostics, significantly reducing the number of false positives and the time-consuming manual curation of candidate variants, allowing a rapid and confident evaluation of putative pathogenic variants. dicast has, however, the potential to be used for other applications such as genome-wide association studies (GWAS) and large-scale aggregation of population variants. Methods Sequencing and optical mapping Illumina paired-end and PacBio CLR sequencing were performed as described in Schöpflin et al., 2022 [32]. Linked-read sequencing by 10x Genomics and Universal Sequencing (TELL-Seq) as well as optical mapping by Bionano Genomics was performed by the respective machine manufacturer according to their best practices. Data preparation and structural variant detection Illumina paired-end WGS Reads were aligned to GRCh38 using bwa-mem (Version 0.7.17) [33]. The resulting BAM files were sorted, and duplicate reads were marked using samtools sort and samtools mdup (Version 1.19.2) [34], respectively. Structural variant detection was performed using five different SV callers: delly (Version 1.1.6), lumpy-smoove (Version 0.2.8), manta (Version 1.6.0), gridss (Version 2.13.2), and cnvnator with bin size 1 kb (Version 0.4.1) [35–39]. All tools were run with default parameters, except indicated otherwise. HLA, decoy or alternate contigs and regions with much higher coverage than expected were excluded from delly and lumpy SV calling (https://github.com/hall-lab/speedseq/blob/master/annotations/exclude.cnvnator_100bp.GRCh38.20170403.bed). PacBio CLR WGS For each sample, subreads in uBAM format were pooled and converted to FASTA format using the tool bam2fastq (Version 1.3.0) [40]. Pooled subreads were aligned to GRCh38 using minimap2 (Version 2.15, parameters -x map-pb -L -MD) [41]. Structural variants were called using the tools svim (Version 1.2.0), sniffles2 (Version 2.0.7), cuteSV (Version 2.0.2), and debreak (Version 1.0.2) [42–45]. All tools were executed with default parameters unless stated otherwise. For a subset of the samples, pooled subreads were de novo assembled with FalconUnzip (Version 0.5) [46]. The assembled diploid contigs were aligned to GRCh38 using minimap2 (Version 2.15, parameters –paf-no-hit -x asm5 –cs –r2k). Finally, SVs were called using the tool svim-asm (Version 1.0.2) with default parameters [47]. 10 x Genomics linked-read WGS Linked-read FASTQ files generated by 10x Genomics were processed with longranger (Version 2.2.2) [48]. Longranger performs sequencing quality control, alignment to GRCh38, as well as detection of SNVs and SVs. For a subset of samples, de novo assembly has been performed using supernova (Version 1.2.0) [49]. Similarly to the long-read data, assembled contigs were aligned to GRCh38 using minimap2 with the same command and structural variants were detected using svim-asm. Universal Sequencing (TELL-Seq) linked-read WGS To process TELL-Seq sequencing data with the longranger pipeline, the barcodes present in the FASTQ files needed to be converted to a longranger compatible format. For that, a tool called ust10x, provided by the sequencing provider Universal Sequencing, has been employed [50]. The tool converts the 18 bp TELL-Seq barcode to a 16 bp 10x compatible barcode while considering a whitelist of 10 x compatible barcodes provided by the user. Subsequently, TELL-Seq FASTQ files were processed analogously to 10x Genomics FASTQ files using the longranger pipeline, supernova and svim-asm. Bionano optical mapping Raw data was processed using the software suite Bionano Solve (Version 1.5.3) [51]. The output obtained from the software consisted of assembled consensus optical maps in CMAP format, aligned consensus maps to a reference genome in XMAP format, as well as structural variant calls in SMAP format. Unless stated otherwise, all subsequent tools and scripts used to further process Bionano data are part of the Solve software suite. SMAP files were converted to VCF format using the script smap_to_vcf_v2.py [52]. The format of the resulting VCF header was adapted using a custom Bash script. In particular, the file format specification line was moved to the correct position within the file, and the length of the reference genome contigs was added to the respective header lines. Next, variants whose start position exceeded their end position were removed, and variants were sorted according to their position using bcftools sort (Version 1.21) [34]. Variant representations were standardized, and duplicate records removed using bcftools norm (Version 1.21, parameters -d all). Population databases As part of the ground truth construction, SV calls from the analyzed samples have been compared to SV calls from publicly available databases. For that, we utilized data from multiple sources, including from the GIAB SV Tier 1 v0.6, gnomAD SV v2.1, and Human Genome Structural Variation (HGSVC) Freeze 3 dataset [15, 16, 53]. Additionally, we included SV calls from a population catalog published by Audano et al. [54]. Variant calls that were only available on GRCh37 were lifted over to GRCh38 using the tool liftover [55]. Variant quality filtering For comparing different sequencing technologies and SV detection methods, a ground truth set of variants needed to be established for each sample. First, a custom Python script was employed to convert VCF files across all methods to a standardized format. Subsequently, all VCF files were filtered to remove low-quality variants as defined by the respective callers. For the methods delly, manta, gridss, cnvnator, svim, sniffles2, cuteSV, debreak all variants which have been assigned the PASS tag were kept. Additionally, variants which have been assigned the IMPRECISE tag were removed. For lumpy, all variants with a quality score below 30 have been removed. The longranger pipelineF, which was applied to 10 x Genomics and TELL-Seq data, generated two structural variant VCF files, one for deletions and one for large SVs, per sample. For both files, variants with the filter tag PASS were kept and those with the tag IMPRECISE were removed. Both files were merged and missing reference contig lengths were added to the VCF header. For the converted optical mapping VCF files, variants whose start position exceeded their end position were removed, and variant representations were standardized, using bcftools norm (Version 1.21). Lastly, duplicate entries defined by SV type, chromosome, start position, and end position were removed using a custom Python script, and VCF files were sorted using bcftools sort. Construction of ground truth The construction of the ground truth can be divided into two major steps. First, for each sample, we identified which calls across its filtered VCF files qualified as high-confidence real SVs based on concordance across methods, technologies, and public catalogs; we refer to these as confirmed variants (criteria defined below). Second, because the same underlying SV is typically reported multiple times across methods, confirmed calls were merged into a single, non-redundant SV set per sample, which constitutes the final ground truth. To identify confirmed SVs we employed multiple criteria, most of which are based on an initial pairwise overlap between all structural variants. For that, we employed the tool truvari (Version 2.2.0, parameters bench -p 0 –sizemin 10 –sizefilt 10 –sizemax 100,000,000 –dup-to-ins) [56]. The default parameters of truvari determine two SVs to be the same if their respective breakpoints are within 500 bps, and they exhibit a size similarity of at least 0.7. Because the nature of optical mapping does not allow for precise pinpointing of SV breakpoints, a different strategy was required to compare OGM VCF files to those generated by other methods. To implement this, each OGM VCF file was divided into 40 separate files, with each file representing a distinct quantile of the breakpoint confidence interval reported by Bionano’s Solve software. The region in which a breakpoint may lie is indicated by this confidence interval. Instead of performing a single comparison of the entire OGM VCF against all other VCF files, 40 pairwise comparisons were carried out—one for each quantile-based OGM VCF file. In each comparison, truvari’s –refdist and –chunksize parameters were set according to the relevant quantile. The resulting overlap sets were then merged in an iterative manner: overlaps involving OGM SVs with narrower confidence intervals were included first, and if the same variant from another method overlapped a higher-interval OGM SV, that subsequent overlap was disregarded. This resulted in a sample-specific list of pairwise overlapping SV calls between all methods and public databases. For Additional file 1: Fig. S1B we report pairwise Jaccard indices between technologies (and between Illumina callers) on a representative sample, S1. The Jaccard index for a pair (A, B) is defined as the number of SV calls supported by both methods divided by the number supported by either. Next, for each sample, a graph G = (V, E) was constructed, where each SV was represented as a vertex v ∈ V, and an edge e ∈ E indicated an overlap between two variants. Finally, connected components C ⊆ V within G were identified, where each connected component C is defined as a maximal set of vertices such that every pair of vertices in C is connected by a path in G. Here, connected components represent SVs which have been detected by multiple methods, each represented by a vertex v ∈ C. Several criteria were employed to select connected components, i.e., variant representations, to be included in the final benchmark set of SVs. First, variants which were supported by at least two methods based on two different sequencing technologies were selected. Furthermore, variants which were supported by at least two different long-read-based methods were included. This criterion was not applied to INVs due to high numbers of spurious INVs in long-read-based callsets. Lastly, variants which were supported by multiple methods of the same technology and were present in any population database were included in the benchmark. This strategy resulted in a set of connected components consisting of multiple variant calls representing the same benchmark SV. Consensus start and end positions were inferred by calculating the sum of absolute differences between each candidate breakpoint and all other breakpoints, selecting the breakpoint pair (left and right) that minimized the total divergence from all other calls. Consensus sizes for insertions were inferred by selecting the most frequently occurring size when a value appeared multiple times in the calls, or when no consensus existed, by choosing the size that minimized the total absolute difference to all other size estimates. Manual curation of ground truth dataset Benchmark SV datasets, comprised of presumed true variants in an individual genome, can contain inaccuracies that arise from the methodologies used in their construction. These errors typically fall into two categories: FPs, which are erroneously included due to shared biases across technologies, and FNs, which are true variants missed because the strengths of certain technologies are underrepresented by multi-technology criteria. To mitigate such errors in our constructed ground truth datasets, we conducted extensive manual curation on a carefully selected subset of variants. We applied the following criteria to identify candidate variants for manual review:Candidate False-Positives: For manual inspection, we selected variants using two criteria: (i) variants present in the benchmark with a dicast score < 0.4, and (ii) for each variant-calling method, the 50 variants with the lowest internal quality scores (below the 0.2 percentile of that method's score distribution). The first criterion accounts for approximately 40% of variants selected for manual curation and the second for approximately 60%.Candidate False-Negatives: Conversely, we selected (i) variants absent from the benchmark with a dicast score > 0.4, and (ii) for each variant-calling method, the 50 variants with the highest internal quality scores (above the 0.8 percentile of that method's score distribution). The first criterion accounts for approximately 40% of variants selected for manual curation and the second for approximately 60%. Candidate False-Positives: For manual inspection, we selected variants using two criteria: (i) variants present in the benchmark with a dicast score < 0.4, and (ii) for each variant-calling method, the 50 variants with the lowest internal quality scores (below the 0.2 percentile of that method's score distribution). The first criterion accounts for approximately 40% of variants selected for manual curation and the second for approximately 60%. Candidate False-Positives: For manual inspection, we selected variants using two criteria: (i) variants present in the benchmark with a dicast score < 0.4, and (ii) for each variant-calling method, the 50 variants with the lowest internal quality scores (below the 0.2 percentile of that method's score distribution). The first criterion accounts for approximately 40% of variants selected for manual curation and the second for approximately 60%. Candidate False-Negatives: Conversely, we selected (i) variants absent from the benchmark with a dicast score > 0.4, and (ii) for each variant-calling method, the 50 variants with the highest internal quality scores (above the 0.8 percentile of that method's score distribution). The first criterion accounts for approximately 40% of variants selected for manual curation and the second for approximately 60%. Candidate False-Negatives: Conversely, we selected (i) variants absent from the benchmark with a dicast score > 0.4, and (ii) for each variant-calling method, the 50 variants with the highest internal quality scores (above the 0.8 percentile of that method's score distribution). The first criterion accounts for approximately 40% of variants selected for manual curation and the second for approximately 60%. Since dicast receives as input variants from the five used SRS-based methods, all variants came from these methods, not only benefiting dicast but also the method used to call this variant originally. The selected variants were manually reviewed by a panel of five human experts using a custom web application. This tool presented visual representations of each variant to facilitate evidence-based decisions on their validity. Based on expert consensus, FP variants were removed from the benchmark set, and FN variants were incorporated. Variant visualization with cuban To support manual curation, we developed a Python-based visualization tool named cuban. The tool generates PNGs that summarize key features of each SV, providing comprehensive context for expert evaluation (see Additional file 1: Fig. S14 for a visual explanation). Information from the alignment of sequencing reads is extracted using pysam [34]. The visualizations include the following tracks: Repetitive elements Repetitive sequences in the human genome can affect read alignment and variant detection. This track displays annotated repeat elements that are near the structural variant. VNTR and STR annotations were obtained from Lu et al. [57]. Other repeat element annotations were obtained from the repeatmasker track of the UCSC Table Browser [58]. Coverage Read coverage across the whole structural variant with padding on both sides. Breakpoints are indicated by vertical black dashed lines. Average chromosomal coverage is indicated by a horizontal red dashed line and was calculated by randomly sampling the average coverage of 1,000 regions of size 1 kb. For the coverage calculation, all reads are considered. A red coverage line displays the coverage when only considering reads with a mapping quality > 20. Insert size outlier distribution Insert size is the distance between two mates of a read-pair. A higher insert size of read pairs around the breakpoints than expected from the sequencing experiment can indicate the presence of an SV. The peaks of the line plot across the entire length of an SV display the number of reads that have an unusually high insert size (larger than 1 kb). Discordant read pairs The distribution of discordant read pairs is assessed across the whole length of the variant. Counts of read pairs in reverse-reverse (RR, orange), reverse-forward (RF, dark blue), forward-forward (FF, green) are displayed. Additionally, read counts where the two mates align to different chromosomes are shown (TX, red). Read alignments In addition to the tracks across the entire length of the SV, cuban visualizes the alignment of reads within windows of 200 bp around each breakpoint. For that, aligned reads are collected with pysam and their CIGAR strings are fed into a matrix where each row corresponds to one read and each column to one nucleotide in the reference genome. The numbers in each cell correspond to different CIGAR string operations, which are subsequently visualized using the following color code:Match/Mismatch: Light grayLow mapping quality (MAPQ): Dark grayDeletion: BlueInsertion: YellowSoft-clip: GreenHard-clip: RedSplit-read: Black stripesRR orientation: Orange stripesRF orientation: Dark blue stripesFF orientation: Green stripes Match/Mismatch: Light gray Match/Mismatch: Light gray Low mapping quality (MAPQ): Dark gray Low mapping quality (MAPQ): Dark gray Deletion: Blue Deletion: Blue Insertion: Yellow Insertion: Yellow Soft-clip: Green Soft-clip: Green Hard-clip: Red Hard-clip: Red Split-read: Black stripes Split-read: Black stripes RR orientation: Orange stripes RR orientation: Orange stripes RF orientation: Dark blue stripes RF orientation: Dark blue stripes FF orientation: Green stripes FF orientation: Green stripes Read connections Split reads contain two segments from the same read that align to different locations in the reference genome. When these segments map to opposite sides of an SV, they can signal the presence of that SV. In such cases, the two aligned segments are connected by a black line in the visualization. Similarly, when two mates align to opposite sides of an SV, they are connected by a red line in the alignment view. Benchmarking SV detection methods and technologies Evaluation of SV detection methods Based on the newly constructed ground truth dataset, a systematic evaluation of SV detection methods and their underlying sequencing or optical mapping technology was performed. A variant was considered confirmed if it overlapped with a benchmark variant using the following criteria: breakpoints within 500 bp of each other and length similarity of at least 70% (for insertions, only breakpoint distance within 200 bp was considered). Overlaps were performed using the Python library bioframe [59]. Performance metrics were then calculated by treating confirmed variants as true positives and unconfirmed variants as false positives, while benchmark variants without matching calls were counted as false negatives. Multiple calls that overlapped with the same benchmark variant were counted as a single TP variant, with the highest quality score being retained among the overlapping calls. Precision and recall were calculated as:Precision = TP / (TP + FP)Precision=TP/(TP+FP)Recall = TP / (TP + FN)Recall=TP/(TP+FN) Evaluation results for each SV caller across multiple samples were aggregated by calculating mean and standard deviation. Evaluation of technologies As a guide for choosing the right sequencing technology for various use cases, we aimed to not only compare SV detection methods but also the sequencing and optical mapping technologies these are based on. To facilitate this comparison, two caller integration strategies per technology were employed. First, a union call set was generated for each technology per sample by pooling all variants called by any of the associated callers. Performance (precision and recall) was then calculated for this union set against the benchmark dataset using the TP, FP, and FN definitions described above. Second, a consensus call set was generated for each technology per sample by retaining only those SVs called by at least 2 different SV detection methods. Again, performance metrics were assessed for this callset against the benchmark. Similar to the evaluation procedure of SV detection methods, performance metrics for each technology were aggregated across multiple samples by calculating mean and standard deviations. Next, for each technology and SV type, performance was assessed for different size categories. The consensus callsets for each technology served as the basis for this analysis. Call and benchmark sets were divided into small (50–100 bp), medium (100 bp-1 kb), large (1–10 kb), as well as very large (> 10 kb) variants and precision and recall were computed as described above. To assess the effect of genomic context on detection performance, the overall performance of a technology (defined by its consensus call set) was compared to its performance when constraining the evaluation to variants overlapping with certain genomic annotations. Genomic annotations used for this analysis were CpG islands, long interspersed nuclear elements (LINEs), SINEs, LTRs, low complexity regions, retroposons, simple repeats, satellite DNA, segmental duplications, short tandem repeats (STRs) as well as variable number tandem repeats (VNTRs). VNTR and STR annotations were obtained from a catalog published by Lu et al. whereas CpG islands were downloaded from the UCSC Table Browser [57, 58]. All other annotations were extracted from the repeatmasker track of the UCSC Table Browser. Overlaps between annotations and variants were calculated using the Python library bioframe. A variant was assigned to an annotation if their corresponding intervals overlapped. For each technology, per-annotation precision and recall were then expressed as log2 fold change relative to that technology's overall precision or recall across all variants (Additional file 1: Fig. S5A). To resolve performance within tandem-repeat regions at finer granularity, we further stratified the STR and VNTR catalogs of Lu et al. STR loci were binned by motif unit length into five classes (2, 3, 4, 5, and ≥ 6 bp). The 1 bp homopolymer class was absent from the catalog and was therefore omitted. VNTR loci were binned by reference-locus length (stop minus start) into five classes (< 200 bp, 200 to 500 bp, 500 bp to 1 kb, 1 to 5 kb, > 5 kb). Each benchmark variant was assigned to its nearest STR or VNTR record using bioframe.closest and retained for the subtype analysis only if its breakpoints fell within 100 bp of that record, consistent with the proximity used for the aggregate STR and VNTR overlap flags. Per-technology precision and recall were then recomputed within each bucket and expressed as log2 fold change relative to the technology's overall baseline. To assess the performance of different technology combinations, the union of their respective consensus call sets was formed. This served as the basis to calculate the number of TPs, FPs, as well as FNs and consequently precision and recall. Development of novel SV detection method Dicast generalizes the concept of consensus calling by implementing a machine learning framework that scores candidate SVs using features derived from genome annotations and sequence alignments. Rather than relying solely on overlapping calls from multiple methods, dicast quantifies the probability that each SV represents a true variant by integrating multiple evidence sources. dicast accepts as input a variable number of VCF files containing candidate structural variants calls from existing methods or candidate variants from a population catalog. Input variants are consolidated into a unified format represented by variant type, chromosome, start, end, as well as SV size. Feature collection Subsequently, variants are annotated with a range of features describing different aspects of a variant. For that, four bins, two around each breakpoint, are defined around each variant. Let BPl denote the position of the left breakpoint and BPr the position of the right breakpoint. Then the coordinates of the four bins are defined as.Bin I: [BPl—52, BPl—2]Bin II: [BPl + 2, BPl + 52]Bin III: [BPr—52, BPr—2]Bin IV: [BPr + 2, BPr + 52] Bin I: [BPl—52, BPl—2] Bin I: [BPl—52, BPl—2] Bin II: [BPl + 2, BPl + 52] Bin II: [BPl + 2, BPl + 52] Bin III: [BPr—52, BPr—2] Bin III: [BPr—52, BPr—2] Bin IV: [BPr + 2, BPr + 52] Bin IV: [BPr + 2, BPr + 52] Since insertions are defined by a single breakpoint, only Bins I and II are considered for this variant type. For each bin, several features derived from the alignment of sequencing reads around the breakpoints are calculated: coverage, insert size, mapping quality, clipped reads, split-reads, and discordant read pairs. Coverage is assessed by calculating the mean read depth within each bin and normalizing it by the average coverage across the corresponding chromosome. To account for local variation, the standard deviation of coverage within each bin is also computed. In addition to the bins near breakpoints, dicast defines four additional bins evenly distributed along the SV body. These internal bins are used exclusively for coverage-based annotations, helping to characterize coverage patterns not only at the breakpoints but across the entire SV. Insert size, defined as the genomic distance between paired-end reads, is measured for each bin by computing the mean insert size of all reads overlapping with a given bin. This mean is normalized by the chromosomal average insert size, and the standard deviation is calculated to capture variability within each bin. Mapping quality is annotated by calculating both the mean and standard deviation of the mapping quality scores of all reads overlapping a given bin. Clipped reads, which are reads that only partially align to the reference genome, can signal potential breakpoints. For each bin, the proportion of clipped reads is calculated as the number of clipped reads divided by the total number of reads overlapping that bin. Split reads, a subset of clipped reads, whose unaligned segments map elsewhere in the genome, are similarly quantified. The proportion of split reads is computed in the same manner as clipped reads. To capture the structural relationship between different regions of a variant, bin connection features are computed. These features quantify the evidence that two distinct bins are connected by abnormal read patterns. For each pairwise combination of bins, two main connection types are assessed:Split-read connections: These involve individual reads that are split across two bins, meaning one segment of the read aligns to one bin and the other segment aligns to the second bin. For each bin pair, the proportion of such split reads (relative to all split reads overlapping either bin) is calculated.Discordant read pair connections: These involve paired-end reads where one mate aligns to one bin and the other mate aligns to the second bin in an unexpected orientation. For each bin pair, the proportion of these discordant pairs is measured. Split-read connections: These involve individual reads that are split across two bins, meaning one segment of the read aligns to one bin and the other segment aligns to the second bin. For each bin pair, the proportion of such split reads (relative to all split reads overlapping either bin) is calculated. Split-read connections: These involve individual reads that are split across two bins, meaning one segment of the read aligns to one bin and the other segment aligns to the second bin. For each bin pair, the proportion of such split reads (relative to all split reads overlapping either bin) is calculated. Discordant read pair connections: These involve paired-end reads where one mate aligns to one bin and the other mate aligns to the second bin in an unexpected orientation. For each bin pair, the proportion of these discordant pairs is measured. Discordant read pair connections: These involve paired-end reads where one mate aligns to one bin and the other mate aligns to the second bin in an unexpected orientation. For each bin pair, the proportion of these discordant pairs is measured. In addition to features derived from aligned sequencing reads, dicast takes the genomic context of a variant into account. The extensive evaluation of SV detection methods and their underlying technologies has revealed that alignment patterns indicating the presence of a TP structural variant can look vastly different depending on their location in the genome. For example, due to difficult read alignments in repetitive regions, positions of clipped and split reads can be shifted. Therefore, each variant is annotated with a list of features derived from reference genome annotations. These features are described in the Methods section Evaluation of technologies. Training The collected features serve as the basis to train an XGBoost model to distinguish TP SVs from FP artifacts. The set of eight individuals (the ninth one serves as a test set) for which ground truths have been created make up the training set. Variants from unfiltered VCF files generated by delly, manta, lumpy, gridss, and cnvnator were annotated with features as described by the process above. The ground truth datasets generated by the integration of multiple sequencing and one optical mapping technology were used to assign labels (TP or FP) to the unfiltered variants. Evaluation For deletions and insertions, dicast was evaluated on the Tier 1 v5.0q stvar benchmark for sample HG002 published by the GIAB Consortium, restricted to its benchmark-region BED. For duplications, evaluated against the held-out S1 sample of our newly created ground-truth dataset. Candidate variants were again called using delly, manta, lumpy, gridss, and cnvnator which were subsequently annotated with the features described above. Additionally, variants from a population catalog published by Ebert et al., were used as input candidate variants for dicast. The population catalog was derived from variants identified in 32 haplotype-resolved human genome assemblies. To avoid data leakage, the test sample HG002 was excluded from the catalog. The trained dicast models were used to score variants for each SV type, respectively. SV-type-specific score cut-offs were used to filter out low-quality, likely FP variants: 0.45 for deletions, 0.30 for insertions, and 0.40 for duplications. The thresholds were chosen per SV type to balance precision and recall on the training cohort, reflecting the differing difficulty of detecting each variant class from short-read alignments. To benchmark dicast against an existing machine-learning-based ensemble approach for SRS-based SV calling, we additionally ran SV2 (v1.4.3.4) on both evaluation samples, HG002 for deletions and insertions and the held-out S1 sample for duplications. For each sample, SV2 was provided with the unfiltered callsets from delly, manta, lumpy, gridss, and cnvnator together with the corresponding BAM and reference genome, and was run with default parameters. Because SV2 additionally requires per-sample SNV input for its genotype-likelihood estimation, we called SNVs from each BAM using freebayes (default parameters) and supplied the resulting VCFs to SV2. SV2 returns a per-variant genotype-quality classifier; for the binary call-set comparison we retained SV2 calls passing its standard PASS filter, and for the precision-recall curves we used SV2's classifier probability as the variant score. For the other SV callers, variants which contained the PASS filter tag, were kept. Subsequently, call sets were overlapped with the respective benchmark datasets. Two variants were called overlapping if their breakpoints were located within 500 bp and their size similarity was > 0.7. The number of TP variants per method was determined by the number of benchmark variants overlapping with a called variant. The number of FP was equal to the number of all variants in the callset which did not overlap with a benchmark variant. FN variants were all benchmark variants which did not overlap with any called variant. To determine precision-recall curves, the respective quality scores of all callers, including SV2's classifier probability, were used. To compare dicast against n-of-N consensus strategies, we pooled the three evaluation settings (DEL and INS on HG002 against GIAB v5, DUP on the held-out S1 sample). For each n from 1 to 5, a variant was retained as a consensus call if at least n of the five Illumina callers (delly, manta, lumpy, gridss, cnvnator; for INS only delly, manta, and gridss as the remaining callers do not produce insertion calls) reported it after caller-specific PASS filtering, and TP and FP counts were obtained against the corresponding benchmark using the same overlap criterion as above. For the score-by-support analysis (Fig. 3C), each benchmark-matched variant was binned by the number of Illumina callers that detected it, and dicast scores were compared between TPs and FPs within each bin. For the rescue analyses (Fig. 3D), we identified per caller the subset of confirmed benchmark variants that (i) were called but received a non-PASS quality flag from that caller ("low-quality") or (ii) were detected by that caller alone among the five Illumina callers ("singletons"), and reported the fraction whose dicast score passed the SV-type-specific threshold. To assess detection of clinically relevant variants, we evaluated the same Illumina callers and dicast on the Genome in a Bottle HG002 CMRG structural variant benchmark v1.00. Benchmark variants were typed from their REPTYPE annotation and sized as |len(REF) − len(ALT)|; we retained events ≥ 50 bp, yielding 101 deletions (REPTYPE CONTRAC, SIMPLEDEL, or SUBSDEL) and 57 simple insertions (SIMPLEINS). Detection was assessed in two modes (Additional file 1: Fig. S6C). In the de novo mode, a benchmark variant was counted as detected by a given Illumina caller if it overlapped a PASS-filtered call of the matching SV type for HG002, and as detected by dicast if it overlapped a dicast prediction passing the SV-type-specific score threshold, using the same breakpoint and length-similarity overlap criterion as above. In the predefined-regions (genotyping) mode, dicast was instead applied directly to the benchmark loci: each CMRG variant was annotated with the reference- and alignment-derived features described above from the HG002 short-read alignments, scored with the corresponding dicast model, and counted as detected if its score exceeded the SV-type-specific threshold. Computational resources To quantify dicast's runtime and memory footprint, we benchmarked it end-to-end against SV2 (v1.4.3.4) on HG002 (30 × paired-end Illumina WGS, GRCh38) on a dual-socket AMD EPYC 7601 host (128 threads, 1 TB RAM, Linux 6.12, Python 3.12). For dicast, we recorded per-stage wall-clock time, CPU time, and peak resident set size (RSS) across the feature collection and prediction stages. dicast is parallelized per SV type per chromosome and was run on the union of delly, manta, lumpy, cnvnator, and gridss candidate calls (70 threads). SV2 does not support multithreading and was run with default parameters on the same BAM, reference, and input VCFs. Wall-clock time, total CPU time, and peak RSS were extracted from its per-stage logs. Evaluation of novel SV detection method in clinical scenarios Short- and long-read processing The experimental procedure and the base pipeline for data processing (alignment and SV calling) of short-read and long-read data has been previously described [27]. The alignment and variant calling closely reflect the data processing described for the ground truth generation but there are some specific methodological differences: No assembly based callers were used and the original overlap computation was performed using reciprocal overlap rather than distance based metrics. We used the initial implementation and created three separate pipelines with minor filtering adjustments generating data-sets that can be used to evaluate dicast: For the first (filtered Illumina) and third pipeline (filtered PacBio and Illumina), we only retained SV calls passing the caller-specific quality filter, i.e., “PASS” variants and for the second pipeline (unfiltered Illumina), we retained all detected variants and scored them using dicast. Variant frequency filtering We compared our call-sets of each pipeline to three catalogs of common variation as well as NCBI nstd186 [16, 53, 54]. From each data-set, we discarded rare variants based on allele frequency or similar metrics (0.01 allele frequency for gnomAD-SV v2.1 and NCBI, “Singletons” for the long-read data-sets). Then we clustered the variants identified in our pipelines with the common variants using a reciprocal overlap threshold of 0.7 and only retained those clusters without any known common variation. Given the phenotypic diversity in our limb malformation cases, we expected unique causative variants and filtered out any variants present in multiple individuals, resulting in a set of singletons. We also investigated shared rare variants between patients in a separate analysis not discussed in this publication. Variant annotation To identify potential functionally relevant variants, we employed an adapted version of the TADA method [28]. We used sets of coding and non-coding annotations tailored to the context of limb malformations. The coding annotations included gene definitions as well as exons, start- and stop-codons, 5’ and 3’ UTRs from ENSEMBL v.110 and developmental-disease associated genes (DECIPHER) [60]. We also generated four additional gene sets specific to the patients’ phenotypes: 1) genes correlated with limb development based on a separate scRNA study 2) genes known to be associated with limb development derived from previous work by the Mundlos AG at the Max-Planck-Institute for Molecular Genetics (MPIMG) 3) phenotype associated genes and 4) differentially expressed genes (DEGs). The phenotype associated genes for each patient are based on the patient’s HPO terms [61]. The DEGs are derived from a patient specific RNA-seq analysis. For non-coding annotations we collected cell-type specific active enhancers determined using cap analysis gene expression (CAGE) from FANTOM5 [62]. Given the lack of limb-specific enhancers in this specific resource, we used the entire collection during annotation. We retrieved the entire set of human candidate regulatory elements (cREs) from ENCODE [63]. While many regulatory elements contained in this set are likely not relevant to limb development, they could indicate regions of potential interest that would warrant further investigation. We also collected a set of experimentally validated enhancers from VISTA [64]. To reflect the limb development-associated regulatory landscape, we collected a set of cREs identified in mouse embryonic limb tissue (14.5 days—ENCODE ID: ENCFF050TOW) and lifted the data over to GRCh38. Given the lack of publicly available limb-specific CTCF sites in humans, we collected ENCODE ChIP-seq derived CTCF narrow peaks from fibroblast samples (ENCFF788LNG). We also included topologically associated domain (TAD) boundaries derived from our Hi-C analysis both as proxies for windows of increased regulatory interactions and as individual annotations. Given the set of coding and non-coding annotations, we defined functionally relevant variants as those hitting either limb-developmental associated coding annotations directly or affecting non-coding regulatory elements in the same TAD environment as disease-relevant genes. This process as well as detailed description of data processing and analysis are available in the referenced dissertation [27]. Manual curation and evaluation In a final step, we performed manual curation using cuban images on the functionally relevant SVs, resulting in a set of candidate variants for each pipeline. For comparison between the candidate sets of the three pipelines we used reciprocal overlap with a threshold of 0.7. The evaluation of dicast with consensus approaches is based on the candidates generated by the second pipeline. In this pipeline multiple SV callers were used and their unfiltered calls scored by dicast followed by a merging process that returns a representative for each cluster of overlapping variants. This resulted in downstream candidates after manual curation with multiple associated dicast scores. For all evaluations we used the highest dicast score for each cluster of merged variants. RNA-seq The fibroblast samples of all patients were cultured in DMEM with 10% FBS, 1% L-glutamine, and 1% pen-strep. RNA extraction from fibroblasts was performed in all 21 samples using the RNeasy mini kit (Qiagen, Hilden, Germany). The Poly(A) mRNA capture was done using the KAPA mRNA HyperPrep Kit (KR1352 -v5.17) and sequencing was performed on a HiSeq4000 (Illumina) using a single technical replicate (PE75, 50 million fragments per sample). We processed the raw sequencing data with a custom snakemake pipeline: First, we aligned the reads to the GRCh38 reference using STAR v.2.7.9a [65]. We filtered for reads with a minimum MAPQ of 5. To identify DEGs for each patient, we conducted a one vs. all analysis with respect to the UCSC hg38 knownGene library using a custom R script. Briefly, we removed any data mapped to sex chromosomes and counted the reads of each patient per exon with summarizeOverlaps. We then applied the DESeq2 (v1.26) function DESeq computing the difference between read counts of a single sample and the remaining cohort [66]. Finally, we normalized the results using vst, computed the logarithmic Fold-Change (Log2FC) per gene, and returned the top 50 genes with the highest Log2FC as the list of DEGs for each patient. Hi-C The PCR amplification (4–8 cycles) of the fibroblast samples was done using NEBNext Ultra II Q5 Master Mix (New England BioLabs, M0544). The PCR purification and size selection was carried out using Agencourt AMPure XP beads (Beckman Coulter, A63881). Libraries were sequenced in 75 bp, or 100 bp paired-end runs on a NovaSeq6000 (Illumina) using between 2 and 5 technical replicates. We conducted the bioinformatic analysis in a custom snakemake pipeline. First, we processed the raw sequencing data using the Juicer pipeline (v.1.6) with bwa-mem (v0.7.17) for aligning the reads to the GRCh38 reference [67]. We then merged the deduplicated and filtered reads across technical replicates. For each patient, we applied the pre function from Juicer-Tools (v1.22.01) to generate Hi-C maps including all mapped reads with MAPQ⩾ 30. In addition, we merged the deduplicated reads across the entire cohort, creating a high-resolution fibroblast Hi-C map. To determine TAD boundaries, we first generated patient-specific VC SQRT normalized contact maps at 25 kb resolution for each chromosome using the Juicer-Tools dump function and then applied TopDom [68]. Rare disease cohorts The experimental procedure of the atrial fibrillation and neuromuscular short-read data were previously described [69–71]. The processing, filtering and annotation was done with the same pipeline as for the limb-malformation cases, with adjustments in the prioritization process and annotations given the patients’ phenotypes. Supplementary Information Additional file 1: Figures S1–S15. Supplementary figures with legends. Contains all supplementary figures referenced in the text.Additional file 2: Tables S1–S4. Supplementary tables with legends. Table S1: SV detection methods used per sample for ground truth construction. Table S2: per-sample de novo assembly statistics. Table S3: overlap graph statistics for ground truth construction. Table S4: known pathogenic SVs used in the clinical diagnostic evaluation of dicast. Additional file 1: Figures S1–S15. Supplementary figures with legends. Contains all supplementary figures referenced in the text. Additional file 2: Tables S1–S4. Supplementary tables with legends. Table S1: SV detection methods used per sample for ground truth construction. Table S2: per-sample de novo assembly statistics. Table S3: overlap graph statistics for ground truth construction. Table S4: known pathogenic SVs used in the clinical diagnostic evaluation of dicast.